Research.

We study how teams of AI agents fail, and share what we find.

ColosseumDo agents collude? TerrariumA test bed for agents ChainHomeThe monitor we’re building

Colosseum · NeurIPS 2026

When agents team up.

A way to test whether AI agents secretly collude. Given a private channel, most models tended to.

Mason Nakamura, Abhinav Kumar, Saswat Das, Sahar Abdelnabi, Saaduddin Mahmud, Ferdinando Fioretto, Shlomo Zilberstein, Eugene Bagdasarian

Terrarium · 2025

A safe place to break things.

An open test bed for teams of AI agents, built to study attacks like rogue agents and poisoned data.

Mason Nakamura, Abhinav Kumar, Saaduddin Mahmud, Sahar Abdelnabi, Shlomo Zilberstein, Eugene Bagdasarian

ChainHome · In progress

Small, fast checkers.

Models that check each action before it runs. The first one caught 31 of 31 fake messages, with no false alarms.

Talk to us.

Researching AI agents? We’d love to compare notes.